Multi-Agent Data Mining (MADM) seeks to harness the general advantages offered by Multi-Agent System (MAS) with respect to the domain of data mining. The research described in this thesis is concerned with Multi-Agent Based Clustering (MABC), thus MADM to support clustering. To investigate the use of MAS technology with respect to data mining, and specifically data clustering, two approaches are proposed in this thesis. The first approach is a multi-agent based approach to clustering using a generic MADM framework whereby a collection of agents with different capabilities are allowed to collaborate to produce a ``best'' set of clusters. The framework supports three clustering paradigms: K-means, K-NN and divisive hierarchical clustering. A number of experiments were conducted using benchmark UCI data sets and designed to demonstrate that the proposed MADM approach can identify a best set of clusters using the following clustering metrics: F-measure, Within Group Average Distance (WGAD) and Between Group Average Distance (BGAD). The results demonstrated that the MADM framework could successfully be used to find a best cluster configuration. The second approach is an extension of the proposed initial MADM framework whereby a ``best'' cluster configuration could be found using cooperation and negotiation among agents. The novel feature of the extended framework is that it adopts a two-phase approach to clustering. Phase one is similar to the established centralised clustering approach (except that it is conducted in a decentralised manner). Phase two comprises a negotiation phase where agents ``swap'' unwanted records so as to improve a cluster configuration. A set of performatives is proposed as part of a negotiation protocol to facilitate intra-agent negotiation. It is this negotiation capability which is the central contribution of the work described in this thesis. An extensive evaluation of the extended framework was conducted using: (i) benchmark UCI data sets and (ii) a welfare benefits data set that provides an exemplar application. Evaluation of the framework clearly demonstrates that, in the majority of cases, this negotiation phase serves to produce a better cluster configuration (in terms of cohesion and separation) than that produced using a simple centralised approach.

The aim of this thesis is to compare several extensions of multimodal logic in terms of their representational succinctness on different classes of models. Succinctness is a natural refinement on the notion of expressivity. Intuitively, given two logics L1 and L2, we say that L1 expresses more succinctly than L2 some properties of a class of models if the L1-formulae expressing the properties in question are significantly shorter than all the equivalent L2-formulae. The precise technical interpretation of "significantly shorter" depends on the case at hand and may mean "exponentially shorter", "nonelementary shorter", etc. This work was motivated by the question of whether public announcement logic (PAL) is exponentially more succinct than multimodal logic (ML) on the class S5 of Kripke models with underlying structures in which all relations are reflexive, symmetric, and transitive.

This thesis presents research work conducted in the field of image mining. More specifically, the work is directed at the employment of image classification techniques to classify images where features of interest are very difficult to distinguish. In this context, three distinct approaches to image classification are proposed. The first is founded on a time series based image representation, whereby each image is defined in terms of histograms that in turn are presented as "time series" curves. A Case Based Reasoning (CBR) mechanism, coupled with a Time Series Analysis (TSA) technique, is then applied to classify new "unseen" images. The second proposed approach uses statistical parameters that are extracted from the images either directly or indirectly. These parameters are then represented in a tabular form from which a classifier can be built on. The third is founded on a tree based representation, whereby a hierarchical decomposition technique is proposed. The images are successively decomposed into smaller segments until each segment describes a uniform set of features. The resulting tree structures allow for the application of weighted frequent sub-graph mining to identify feature vectors representing each image. A standard classifier generator is then applied to this feature vector representation to produce the desired classifier. The presented evaluation, applied to all three approaches, is directed at the classification of retinal colour fundus images; the aim is to screen for an eye condition known as Age-related Macular Degeneration (AMD). Of all the approaches considered in this thesis, the tree based representation coupled with weighted frequent sub-graph mining produced the best performance. The evaluation also indicated that a sound foundation has been established for future potential AMD screening programmes.

Trust and reputation are key issues in the multi-agent systems domain. As in human societies, software agents must interact with other agents in settings where there is the possibility that they can be exploited. This suggests the need for theoretical and computational models of trust and reputation that can be used by software agents, and accordingly, much research has investigated this issue. The first part of this thesis investigates the conjecture that agents who make decisions in scenarios where trust is important can benefit from the use of a social structure, representing the social relationships that exist between agents. To this end, we present techniques that can be used by agents to initially build and then progressively update such a structure in the light of experience. As the agents interact with other agents they gather information about interactions and relationships in order to build the network of agents and to better understand their social environment. We also show empirical evidence that a trust model enhanced with a social structure representation, used to gather additional information to select trustworthy agents for an agent’s interactions, can improve the trust model’s performance. In the second part of this thesis, we concentrate on the context of coalition formation. Coalition stability is a crucial issue. Stability is the motivation of an agent’s refusal to break from the original coalition and form a new one. Lack of trust in some of the coalition members could induce one agent to leave the coalition. Therefore we address the current model’s limitation by introducing an abstract framework that allows agents to form distrust-free coalitions. Moreover we present measures to evaluate the trustworthiness of the agent with respect to the whole society or to a particular coalition. We also describe a way to combine the trust and distrust relationships to form coalitions which are still distrust-free.

The formal analysis of computational processes is by now a well-established field. However, in practical scenarios, the problem of how we can formally verify interactions with humans still remains. This thesis is concerned with addressing this problem through the use of the Brahms language. Our overall goal is to provide formal verification techniques for human-agent teamwork, particularly astronaut-robot teamwork on future space missions and human-robot interactions in health-care scenarios modelled in Brahms.

If autonomous agents operating with other agents in open systems are to fulfil their goals and design objectives, the need to discuss and agree upon plans of action is imperative. In this thesis I present work covering both theoretical research and practical development related to the use of argumentation-based dialogues as a way to coordinate actions in multi-agent planning scenarios. The necessity of coordination in multi-agent systems requires the development of mechanisms to propose, modify, share, monitor, and argue about plans. In this thesis I present an argumentation scheme to propose multi-agent plans and associated critical questions to critique the proposal. Such a detailed consideration of multi-agent plan composition contains the right characteristics to enable the justification of plans. This research builds upon research on practical reasoning for action proposals and considers multi-agent plan proposals where plans require several agents for their execution. A dialogue game protocol is also presented which is based on proposal framework. The protocol allows agents to engage in dialogues to agree on and modify plans based on persuasion and deliberation protocols. The detail encompassed by the argumentation scheme and critical questions means that there is a large number of critical questions, and so dialogues may be very lengthy. To overcome this issue, I investigated the issue of strategies for use with this dialogue game in terms of the different possible orderings in which critiques can be posed. The thesis presents an implementation that realises the theoretical framework in terms of a agents engaging in simulated dialogues to share and agree on a plan. The experiments allow us to investigate the effects of such strategies in terms of the number of questions issued to reach an agreement. Overall, the framework presented in this thesis allow agents to engage in dialogues over cooperative plan proposals in a structured way using well-founded argumentative principles.

This thesis presents data mining research work undertaken in the context of identifying correlations between 3D surfaces. More specifcally, this research is directed at predicting distortions (referred to as springback) in sheet metal forming. The main objective was to identify a mechanism that 'best' serves to both capture effectively 3D geometrical information while at the same time allowing for the generation of effective predictors (classifiers). To this end, three distinct 3D surface representation techniques are proposed based on three different concepts. The first technique, the Local Geometry Matrices (LGM) representation, is founded on the idea of Local Binary Patterns (LBPs), as used with respect to image texture analysis, whereby surfaces are defined in terms of local neighbourhoods surrounding individual points in a 3D surface. The second technique, the Local Distance Measure (LDM) representation, is influenced by the observation that springback is greater further from edges and corners, consequently surfaces are defined in terms of distance to the nearest edge or corner. The third technique, the Point Series (PS) representation, is founded on the idea of using a spatial 'linearisation' with which to represent surfaces in terms of point series curves. The thesis describes and discusses each of these in detail including, in each case, the theoretical underpinning supporting each representation. A full evaluation of each of the representations is also presented. As will become apparent, the PS technique was found to be the most effective. The presented evaluation was directed at predicting springback, in the context of the Asymmetric Incremental Sheet Forming (AISF) manufacturing process, in such a way that an enhanced version of the desired 3D surface can be proposed intended to minimise the effect of springback. For the evaluation two flat-topped, square-based, pyramid shapes were used. Each pyramid had been manufactured twice using Steel and twice using Titanium. In addition this thesis presents some ideas on how the springback prediction mechanism can be incorporated into an 'intelligent process model'. The evaluation of this model, by manufacturing corrected shapes, established that a sound prediction framework, incorporating the 3D surface representation techniques espoused in this thesis coupled with a compatible classification technique, had been established.

Optimization techniques are widely applied in various engineering areas, such as modeling, identification, optimization, prediction, forecasting and control of complex systems. This thesis presents the novel optimization methods that are used to control Photovoltaic (PV) generation systems. PV power systems are electrical power systems energized by PV modules or cells. This thesis starts with the introduction of PV modeling methods, on which our research is based. Parameter estimation is used to extract the parameters of the PV models characterizing the utilized PV devices. To improve efficiency and accuracy, we proposed sequential Cuckoo Search (CS) and Parallel Particle Swarm Optimization (PPSO) methods to extract the parameters for different PV electrical models. Simulation results show the CS has a faster convergence rate than the traditional Genetic Algorithm (GA), Pattern Search (PS) and Particle Swarm Optimization (PSO) in sequential processing. The PPSO, with an accurate estimation capability, can reduce at least 50% of the elapsed time for an Intel i7 quad-core processor. A major challenge in the utilization of PV generation is posed by its non linear Current-Voltage (I-V ) relations, which result in the unique Maximum Power Point (MPP) varying with different atmospheric conditions. Maximum Power Point Tracking (MPPT) is a technique employed to gain maximum power available from PV devices. It tracks operating voltage corresponding to the MPP and constrains the operating point at the MPP. A novel model-based two-stage MPPT strategy is proposed in this thesis to combine the offline maximum power point estimation using the Weightless Swarm Algorithm (WSA) with an online Adaptive Perturb & Observe (APO) method. In addition, an Approximate Single Diode Model (ASDM) is developed for the fast evaluations of the output power. The feasibility of the proposed method is verified in an MPPT system implemented with a Single-Ended Primary-Inductor Converter (SEPIC). Simulation results show the proposed MPPT method is capable of locating the operating point to the MPP under various environmental conditions.

Agent-based decision-making usually relies upon game theoretic principles that are 'rational' i.e. decision-making is purely mathematical based on utilities such as the wealth of an agent. In the context of public goods games, such reasoning can often lead to non-optimal, destructive outcomes for both individuals and the total system, as shown in many scenarios from game theory. This thesis considers how the use of \textit{emotions} can impact upon decision-making and social interactions amongst agents in the iterated Prisoner's Dilemma game by modelling emotions in a functional manner. The background to the thesis is first presented in chapters 2 and 3 where the argument for emotions being included in agent-based decision-making, and evidence to support this proposition, is outlined. Various philosophical issues are also considered such as: do emotions directly motivate an agent's intentional behaviour and, is an agent's decision-making still rational if emotions are used? The framework developed to allow for modelling of emotions in agents is then discussed in chapters 4 and 5 where major psychological models of emotion and computational implementations thereof are discussed. Finally chapters 6 to 8 present extensive investigations into how the emotions modelled using the framework affect social interactions amongst agents in the context described above. As of yet, this topic has been relatively unexplored by computer science and there is space for novel, interesting contributions to be made, these contributions are outlined below. In chapter 6 the emotions of \textit{anger} and \textit{gratitude} are modelled and their effects upon social interactions are analysed. In particular, I look at whether agents endowed with these emotions offer any improvement upon the success of agents using with the ``tit-for-tat'' strategy when playing against other leading strategies from Axelrod's famous computer tournament. How these emotions affect rates of cooperation/defection and the fairness of individual scores is considered along with why they do so. This investigation is furthered in chapter 7 where \textit{admiration} is modelled and an investigation is performed into what emotional characters are selected for under different initial conditions and why. This examination provides a discussion regarding what emotional social norms emerge in a population when agents admire the individual success of others. Two salient questions are asked: is it is the case that emotional characters which promote the total wealth of the system are selected for as an emergent property and, do different initial conditions affect the emotional characteristics selected for?. Finally, chapter 8 extends chapter 7 by modelling \textit{hope} and enquires as to how particular emotional character populations (after a complete social norm has been established) deal with destabilisation of cooperation cycles due to periodic defection. The performance of agents endowed with differing emotional characters are again tested under different initial conditions and specific behavioural features of particular emotional characters are considered. In doing this I comment upon how different emotional characters deal with periodic defection and what the best approach is both in context of an agent's individual score and the total score of the system.

Algorithmic mechanism design studies the allocation of resources to selfish agents, who might behave strategically to maximize their own utilities. This thesis studies these incentive issues arsing from four different settings, that are motivated by real- life applications. We model the settings and problems by appropriately extending or generalizing classical economic models. After that we systematically analyze the auction design problems by using methods from both economic theory and computer science. The first problem is the auction design problem for selling online rich media ad- vertisement. In this market, multiple advertisers compete for a set of slots that are arranged in a line, such as a banner on a website. Each buyer desires a particular num- ber of consecutive slots and has a private per-click valuation while each slot is associated with a quality factor. Our goal is to maximize the auctioneer’s expected revenue given buyers’ consecutive demand. This is motivated by modeling buyers who may require these to display a large size ad. Three major pricing mechanisms, the Bayesian pric- ing model, the maximum revenue market equilibrium model and an envy-free solution model are studied in this setting. The second setting is for fund-raising scenarios, where a revenue target is usually specified. We are interested in designing truthful auctions that maximize the probability to achieve this revenue target, rather than in maximizing the expected revenue. We study this topic from the perspective of Bayesian auction design in digital good auctions. We present an algorithm to find the optimal truthful auction for two buyers with independent valuations and show the problem is NP-hard when the number of buyers is arbitrary or the distributions are correlated. We also investigate simple auctions in this setting and provide approximately optimal solutions. Third, we study double auction market design where the trading broker wants to maximize its total revenue by buying low from the sellers and selling high to the buyers in a Bayesian setting. For single-parameter setting, we develop a maximum mechanism for the market maker to maximize its own revenue. For the more general case where each seller’s product may be different, we consider a number of various settings in terms of constraints on supplies and demands. For each of them, we develop a polynomial time computable truthful mechanism for the market maker to achieve a revenue at least a constant factor times the revenue of any other truthful mechanism. Finally, we study the inefficiency of mixed equilibria of all-pay auctions in three different environments – combinatorial, multi-unit and single-item auctions. First, we consider item-bidding combinatorial auctions where m all-pay auctions run in parallel, one for each good. For fractionally subadditive valuations, we strengthen the upper bound by proving some structural properties of mixed Nash equilibria. Next, we design an all-pay mechanism with a randomized allocation rule for the multi-unit auction, which admits a unique, approximately efficient, pure Nash equilibrium. Finally, we analyze single-item all-pay auctions motivated by their connection to crowdsourcing contests and show tight bounds on the PoA of social welfare, revenue and maximum bid.